feat: unify attention backend with multi-dim mask support

- Add attention() functional entry delegating to active backend
- GQA/MLA forward calls attention() instead of inline cache/SDPA
- CUDA kernels support 2D/3D/4D mask via mask_h_stride field
- CudaBackend.fwd_decode builds 2D padding mask for mixed seq_lens
- KVCache.max_len precomputed in bind_tasks to avoid GPU sync
- batch==1 decode short-circuits mask=None
- Split tests into conftest, test_backend, test_backend_equivalence, test_kernel_mask
- 440 tests pass, L20 decode 1.44-1.60x speedup vs torch native

csrc/kernels/attn_prefill_split_q.cuh

@@ -64,7 +64,7 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
 
     // KV: stride-based base
     int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
-    int mask_batch_base = batch * p.mask_b_stride;
+    int mask_batch_base = batch * p.mask_b_stride + q_head * p.mask_h_stride;
     int tiles   = (p.kv_len + P_BC - 1) / P_BC;
     int tt      = G * ROWS;
     int lid     = row * G + gpos;